public void Forwerd(R1dArray input, R2dArray w, out R1dArray u, out R1dArray o, Activator act)
{
u = new R1dArray(w.Height, input.Batch);
o = new R1dArray(w.Height, input.Batch);
var ipt = input << 1;
var _u = u;
var _o = (RNdArray)o;
if (Option_F_01 != null)
{
Option_F_01.Startup();
using (Cloo.ComputeBuffer <float> __ipt = Option_F_01.ConvertBuffer(Cloo.ComputeMemoryFlags.ReadOnly, ipt.Data))
using (Cloo.ComputeBuffer <float> __w = Option_F_01.ConvertBuffer(Cloo.ComputeMemoryFlags.ReadOnly, w.Data))
using (Cloo.ComputeBuffer <float> __u = Option_F_01.ConvertBuffer(Cloo.ComputeMemoryFlags.WriteOnly, u.Data))
{
Option_F_01.SetParameter(__ipt);
Option_F_01.SetParameter(__u);
Option_F_01.SetParameter(__w);
Option_F_01.SetParameter(w.Width, Gpgpu.ProgramOption.ValueMode.INT);
Option_F_01.SetParameter(w.Height, Gpgpu.ProgramOption.ValueMode.INT);
Option_F_01.Execute(ipt.Batch, w.Height);
Option_F_01.ReadBuffer(__u, ref u.Data);
}
}
else
{
Parallel.For(0, ipt.Batch, b =>
{
Parallel.For(0, w.Height, j =>
{
for (int i = 0; i < w.Width; i++)
{
_u[j, b] += w[i, j] * ipt[i, b];
}
});
});
}
if (act != null)
{
act.Activation(u, ref _o);
}
else
{
Parallel.For(0, u.TotalLength, i =>
{
_o.Data[i] = _u.Data[i];
});
}
}